llama.cpp patch reduces MTP buffer overhead for AMD GPUs

A patch reducing MTP buffer overhead increased the usable context length for Qwen 27B from 64K to 149K on a 16GB+12GB GPU setup. The optimization improves memory efficiency for both ROCm and Vulkan backends.